Messages in this thread | | | Date | Tue, 6 Jul 2010 21:47:34 -0400 | From | Christoph Hellwig <> | Subject | Re: [regression] Crash in wb_clear_pending() |
| |
On Tue, Jul 06, 2010 at 08:47:33AM +0200, Ingo Molnar wrote: > Yeah, i agree - especially since the smaller patch is still pretty large (not > a oneliner), plus it does not appear that the precise failure mode is fully > understood either.
The two failure modes we saw are very precicely understood:
- access to ->state after it's been freed - access to ->state after the caller which has the bdi_work struct on stack has been woken up.
But I seriously fear that there are even more dragons hidden in this maze of bit wakeups and rcu, so replacing it with a proper completion seems by far safer to me.
| |